CSE 527Computational BiologyAutumn 2005

Lectures ~15-16Gene Prediction

2

Some References A great online bib

http://www.nslij-genetics.org/gene/ A good intro survey

JM Claverie (1997) "Computational methods forthe identification of genes in vertebrate genomicsequences” Human Molecular Genetics,6(10)(review issue): 1735-1744.

A gene finding bake-off M Burset, R Guigo (1996), "Evaluation of gene

structure prediction programs", Genomics, 34(3):353-367.

3

Motivation Sequence data flooding into Genbank What does it mean?

protein genes, RNA genes, mitochondria,chloroplast, regulation, replication, structure,repeats, transposons, unknown stuff, …

4

Protein Coding Nuclear DNA Focus of next 2 lectures Goal: Automated annotation of new

sequence data State of the Art:

predictions ~ 60% similar to real proteins ~80% if database similarity used lab verification still needed, still expensive

5

Biological Basics Central Dogma: DNA transcription RNA translation Protein

Codons: 3 bases code one amino acid Start codon Stop codons 3’, 5’ Untranslated Regions (UTR’s)

6

The Genetic Code

7

Translation: mRNA → Protein

Watson, Gilman, Witkowski, & Zoller, 1992

8

Ribosomes

Watson, Gilman, Witkowski, & Zoller, 1992

9

Idea #1: Find Long ORF’s Reading frame: which of the 3 possible

sequences of triples does the ribosomeread?

Open Reading Frame: No stop codons In random DNA

average ORF = 64/3 = 21 triplets 300bp ORF once per 36kbp per strand

But average protein ~ 1000bp

10

Idea #2: Codon Frequency In random DNA

Leucine : Alanine : Tryptophan = 6 : 4 : 1 But in real protein, ratios ~ 6.9 : 6.5 : 1 So, coding DNA is not random Even more: synonym usage is biased

(in a species dependant way)examples known with 90% AT 3rd base

11

Recognizing Codon Bias Assume

Codon usage i.i.d.; abc with freq. f(abc) a1a2a3a4…a3n+2 is coding, unknown frame

Calculate p1 = f(a1a2a3)f(a4a5a6)…f(a3n-2a3n-1a3n) p2 = f(a2a3a4)f(a5a6a7)…f(a3n-1a3n a3n+1) p3 = f(a3a4a5)f(a6a7a8)…f(a3n a3n+1a3n+2) Pi = pi / (p1+p1+p3)

More generally: k-th order Markov model k=5 or 6 is typical

12

Codon Usage in Φx174

Staden & McLachlan, NAR 10, 1 1982, 141-156

13

Promoters, etc. In prokaryotes, most DNA coding

E.g. ~ 70% in H. influenzae Long ORFs + codon stats do well But obviously won’t be perfect

short genes 5’ & 3’ UTR’s

Can improve by modeling promoters & othersignals e.g. via WMM or higher-order Markov models

14

As in prokaryotes (but maybe more variable) promoters start/stop transcription start/stop translation

New Features: polyA site/tail introns, exons, splicing branch point signal alternative splicing

Eukaryotes

5’ 3’exon intron exon intron

AG/GT yyy..AG/G AG/GT donor acceptor donor

15

Characteristics of human genes(Nature, 2/2001, Table 21)

* 1,804 selected RefSeq entries were those with full-length unambiguous alignment to finished sequence

Selected RefSeq entries (1,804)*27 kb14 kbGenomic extent

447 aa367 aa(CDS)

Selected RefSeq entries (1,804)*1340bp1,100 bpCoding seq

Confirmed by mRNA or EST on chromo 22 (463)300 bp240 bp5' UTR

Confirmed by mRNA or EST on chromo 22 (689)770 bp400 bp3' UTR

RefSeq alignments to finished sequence (27,238 introns)3,365 bp1,023 bpIntrons

RefSeq alignments to finished sequence (3,501 genes)8.87Exon number

RefSeq alignments to draft genome sequence, withconfirmed intron boundaries (43,317 exons)

145 bp122 bpInternal exon

Sample (size)MeanMedian

16

Big Genes Many genes are over 100 kb long, Max known: dystrophin gene (DMD), 2.4 Mb. The variation in the size distribution of coding

sequences and exons is less extreme,although there are remarkable outliers. The titin gene has the longest currently known

coding sequence at 80,780 bp; it also has thelargest number of exons (178) and longest singleexon (17,106 bp).

RNApol rate: 2.5 kb/min

17

Nature 2/2001

18

Figure 36 GC content. a,Distribution of GC content ingenes and in the genome. For9,315 known genes mapped tothedraft genome sequence, the localGC content was calculated in awindow covering either the wholealignment or 20,000 bp centredaround the midpoint of thealignment, whichever was larger.Ns in the sequence were notcounted. GC content for thegenome was calculated foradjacent nonoverlapping 20,000-bp windows across the sequence.Both the gene and genomedistributions have beennormalized to sum to one.

b, Gene density as a functionof GC content, obtained bytaking the ratio of the data ina. Values are less accurate athigher GC levels because thedenominator is small. c,Dependence of mean exonand intron lengths on GCcontent. For exons andintrons, the local GC contentwas derived from alignmentsto finished sequence only,and were calculated fromwindows covering the featureor 10,000 bp centred on thefeature, whichever waslarger.

Nature 2/2001

19

A Case Study -- Genscan C Burge, S Karlin (1997), "Prediction of

complete gene structures in humangenomic DNA", Journal of MolecularBiology , 268: 78-94.

20

Training Data 238 multi-exon genes 142 single-exon genes total of 1492 exons total of 1254 introns total of 2.5 Mb

NO alternate splicing, none > 30kb, ...

21

Performance Comparison

After Burge&Karlin, Table 1. Sensitivity, Sn = TP/AP; Specificity, Sp = TP/PP

Program Sn Sp Sn Sp Avg. ME WE

GENSCAN 0.93 0.93 0.78 0.81 0.80 0.09 0.05

FGENEH 0.77 0.88 0.61 0.64 0.64 0.15 0.12

GeneID 0.63 0.81 0.44 0.46 0.45 0.28 0.24

Genie 0.76 0.77 0.55 0.48 0.51 0.17 0.33

GenLang 0.72 0.79 0.51 0.52 0.52 0.21 0.22

GeneParser2 0.66 0.79 0.35 0.40 0.37 0.34 0.17

GRAIL2 0.72 0.87 0.36 0.43 0.40 0.25 0.11

SORFIND 0.71 0.85 0.42 0.47 0.45 0.24 0.14

Xpound 0.61 0.87 0.15 0.18 0.17 0.33 0.13

GeneID‡ 0.91 0.91 0.73 0.70 0.71 0.07 0.13

GeneParser3 0.86 0.91 0.56 0.58 0.57 0.14 0.09

per exonper nuc.

Accuracy

22

Generalized HiddenMarkov Models π: Initial state distribution aij: Transition probabilities One submodel per state Outputs are strings gen’ed by submodel Given length L

Pick start state q1 (~π) While

Pick string si of length di = |si| ~ submodel for qi Pick next state qi+1 (~aij)

Output s1s2…!

di< L"

23

24

GHMM Structure

25

Leng

th D

istrib

utio

nsAT-rich avg: 2069CG-rich avg: 518

(a) Introns (b) Initialexons

(c) Internal exons

(d) Terminal exons

26

Effect of G+C ContentGroup I II III IV

C ‡ G% range <43 43-51 51-57 >57

Number of genes 65 115 99 101

Est. proportion single-exon genes 0.16 0.19 0.23 0.16

Codelen: single-exon genes (bp) 1130 1251 1304 1137

Codelen: multi-exon genes (bp) 902 908 1118 1165

Introns per multi-exon gene 5.1 4.9 5.5 5.6

Mean intron length (bp) 2069 1086 801 518

Est. mean transcript length (bp) 10866 6504 5781 4833

Isochore L1+L2 H1+H2 H3 H3

DNA amount in genome (Mb) 2074 1054 102 68

Estimated gene number 22100 24700 9100 9100

Est. mean intergenic length 83000 36000 5400 2600

Initial probabilities:

Intergenic (N) 0.892 0.867 0.54 0.418

Intron (I+, I- ) 0.095 0.103 0.338 0.388

5' Untranslated region (F+, F-) 0.008 0.018 0.077 0.122

3' Untranslated region (T+, T-) 0.005 0.011 0.045 0.072

27

Submodels 5’ UTR

L ~ geometric(769 bp), s ~ MM(5) 3’ UTR

L ~ geometric(457 bp), s ~ MM(5) Intergenic

L ~ geometric(GC-dependent), s ~ MM(5) Introns

L ~ geometric(GC-dependent), s ~ MM(5)

28

Submodel: Exons Inhomogenious 3-periodic 5th order

Markov models Separate models for low GC (<43%),

high GC Track “phase” of exons, i.e. reading

frame.

29

Signal Models I: WMM’s Polyadenylation

6 bp, consensus AATAAA Translation Start

12 bp, starting 6 bp before start codon Translation stop

A stop codon, then 3 bp WMM

30

Signal Models II: more WMM’s Promoter

70% TATA 15 bp TATA WMM s ~ null, L ~ Unif(14-20) 8 bp cap signal WMM

30% TATA-less 40 bp null

31

Signal Models III: W/WAM’s Acceptor Splice Site (3’ end of intron)

[-20..+3] relative to splice site modeled by “1storder weight array model”

Branch point & polypyrimidine tract Hard. Even weak consensus like YYRAY found in

[-40..-21] in only 30% of training “Windowed WAM”: 2nd order WAM, but averaged

over 5 preceding positions“captures weak but detectable tendency toward YYY tripletsand certain branch point related triplets like TGA, TAA, …”

32

intron

exon 5’

exon

What's in the Primary Sequence?

33

Signal Models IV: MaximumDependence Decomposition Donor splice sites (5’ end of intron)

show dependencies between non-adjacent positions, e.g. poor match atone end compensated by strong matchat other end, 6 bp away

Model is basically a decision tree Uses χ2 test to quantitate dependence

34

χ2 test for independencei Con j: -3 -2 -1 +3 +4 +5 +6 Sum

-3 c/a --- 61.8* 14.9 5.8 20.2* 11.2 18.0* 131.8*

-2 A 115.6* --- 40.5* 20.3* 57.5* 59.7* 42.9* 336.5*

-1 G 15.4 82.8* --- 13.0 61.5* 41.4* 96.6* 310.8*

+3 a/g 8.6 17.5* 13.1 --- 19.3* 1.8 0.1 60.5*

+4 A 21.8* 56.0* 62.1* 64.1* --- 56.8* 0.2 260.9*

+5 G 11.6 60.1* 41.9* 93.6* 146.6* --- 33.6* 387.3*

+6 t 22.2* 40.7* 103.8* 26.5* 17.8* 32.6* --- 243.6*

* means chi-squared p-value < .001

!

" 2 =(observed

i#expcted

i)2

expectedii

$

“expected” means expectedassuming independence

35

36

GHMM Structure

37

Summary of Burge & Karlin Coding DNA & control signals

nonrandom Weight matrices, WAMs, etc. for controls Codon frequency, etc. for coding

GHMM nice for overall architecture Careful attention to small details pays

38

Problems with BK training set 1 gene per sequence Annotation errors Single exon genes over-represented? Highly expressed genes over-represented? Moderate sized genes over-represented?

(none > 30 kb) … Similar problems with other training sets, too

39

Problems with all methods Pseudo genes Short ORFs Sequencing errors Non-coding RNA genes & spliced UTR’s Overlapping genes Alternative splicing/polyadenylation Hard to find novel stuff -- not in training Species-specific weirdness -- spliced leaders,

polycistronic transcripts, RNA editing…

40

Other ideas Database search - does gene you’re

predicting look anything like a knownprotein?

Comparative genomics - what does thisregion look like in related organisms?